TY - GEN
T1 - Constant-Time Gaussian Filtering for Acceleration of Structure Similarity
AU - Sasaki, Tomohiro
AU - Fukushima, Norishige
AU - Maeda, Yoshihiro
AU - Sugimoto, Kenjiro
AU - Kamata, Sei Ichiro
N1 - Funding Information:
This work was supported by JSPS KAKENHI JP17H01764, 18K18076, 18K19813, 19K24368.
Publisher Copyright:
© 2020 IEEE.
PY - 2020/3/6
Y1 - 2020/3/6
N2 - In this paper, we propose an acceleration method of structural similarity (SSIM) and its multi-scaled version, called MS-SSIM. The calculation process of SSIM and MS-SSIM includes multiple Gaussian filters, and the cost of the filter is dominant for the entire process; thus, to accelerate SSIM/MS-SSIM, we replace Gaussian filtering using convolution with sliding DCT. Gaussian filter based on sliding DCT is faster than the usual convolution method. Besides, its computational complexity does not depend on the filter window length. Also, naive implementations of SSIM and MS-SSIM scan image many times for the pixel-wise operation; however, these operations can be incorporated into Gaussian filtering. Thus, we optimize the processing pipeline to achieve high cache-efficiency. As a result, the proposed SSIM computation was accelerated by 6.36 times and MS-SSIM by 8.11 times faster than the conventional approach.
AB - In this paper, we propose an acceleration method of structural similarity (SSIM) and its multi-scaled version, called MS-SSIM. The calculation process of SSIM and MS-SSIM includes multiple Gaussian filters, and the cost of the filter is dominant for the entire process; thus, to accelerate SSIM/MS-SSIM, we replace Gaussian filtering using convolution with sliding DCT. Gaussian filter based on sliding DCT is faster than the usual convolution method. Besides, its computational complexity does not depend on the filter window length. Also, naive implementations of SSIM and MS-SSIM scan image many times for the pixel-wise operation; however, these operations can be incorporated into Gaussian filtering. Thus, we optimize the processing pipeline to achieve high cache-efficiency. As a result, the proposed SSIM computation was accelerated by 6.36 times and MS-SSIM by 8.11 times faster than the conventional approach.
KW - SSIM
KW - acceleration
KW - constant-time Gaussian filtering
KW - fast image quality assessment
KW - sliding DCT
UR - http://www.scopus.com/inward/record.url?scp=85099473317&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85099473317&partnerID=8YFLogxK
U2 - 10.1109/ICIP48927.2020.9367337
DO - 10.1109/ICIP48927.2020.9367337
M3 - Conference contribution
AN - SCOPUS:85099473317
T3 - Proceedings of International Conference on Image Processing and Robotics, ICIPRoB 2020
BT - Proceedings of International Conference on Image Processing and Robotics, ICIPRoB 2020
A2 - Sudantha, B. H.
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 1st International Conference on Image Processing and Robotics, ICIPRoB 2020
Y2 - 6 March 2020 through 8 March 2020
ER -